Distributed Deep Reinforcement Learning: Learn how to play Atari games in 21 minutes

By Igor Adamski, Igor Adamski, Robert Adamski, Tomasz Grel, Adam Jedrych, Kamil Kaczmarek, Henryk Michalewski. 2018

37 citations recorded by the source corpus.

We present a study in Distributed Deep Reinforcement Learning (DDRL) focused on scalability of a state-of-the-art Deep Reinforcement Learning algorithm known as Batch Asynchronous Advantage ActorCritic (BA3C). We show that using the Adam optimization algorithm with a batch size of up to 2048 is a viable choice for carrying out large scale machine learning computations. This, combined with careful reexamination of the optimizer's hyperparameters, using synchronous training on the node level (while keeping the local, single node part of the algorithm asynchronous) and minimizing the memory footprint of the model, allowed us to achieve linear scaling for up to 64 CPU nodes. This corresponds to a training time of 21 minutes on 768 CPU cores, as opposed to 10 hours when using a single node with 24 cores achieved by a baseline single-node implementation.

Subfields: Artificial Intelligence

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

Citations in the corpus

IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures (2018, neutral): “, 2017), distributed BA3C (Adamski et al., 2018) and Ape-X (Horgan et al.”
Challenges of Real-World Reinforcement Learning (2019, background): “In addition, since there is often only one instance of the system, approaches that instantiate hundreds or thousands of environments to collect more data for distributed training are usually not compatible with this setup (Horgan et al., 2018; Espeholt et al., 2018; Adamski et al., 2018).”
Challenges of real-world reinforcement learning: definitions, benchmarks and analysis (2021, background): “In addition, since there is often only one instance of the system, approaches that instantiate hundreds or thousands of environments to accelerate training through distributed training (Horgan et al. 2018; Espeholt et al. 2018b; Adamski et al. 2018) nevertheless require as much data and are rarel”
An empirical investigation of the challenges of real-world reinforcement learning (2020, background): “In addition, since there is often only one instance of the system, approaches that instantiate hundreds or thousands of environments to accelerate training through distributed training [Horgan et al. 2018; Espeholt et al. 2018b; Adamski et al. 2018] nevertheless require as much data and are rarely c”
Stochastic Activation Actor Critic Methods (2019, methodology): “Recently, batch A3C-CNN was developed for faster training and eﬃcient utilization of GPUs [1].”
GPU-Accelerated Robotic Simulation for Distributed Reinforcement Learning (2018, methodology): “[8] furthered scaled up A3C training by using larger batchsizes and a well-tuned Adam Optimizer.”